ECHO: A Participatory Framework for Bias-Anchored AI Harm Anticipation

Nicoleta Tantalaki; Sophia Vei; Athena Vakali

cs.CYNov 27, 2025

ECHO: A Participatory Framework for Bias-Anchored AI Harm Anticipation

Authors: Nicoleta Tantalaki, Sophia Vei, Athena Vakali

Organizations: Aristotle University, Thessaloniki, Greece

Abstract

Artificial Intelligence (AI) systems increasingly shape consequential decisions, creating value but also potential harms for individuals, social groups, and society. This has prompted calls for proactive approaches that anticipate harms early in the AI lifecycle. Although prior research identifies AI biases as sources of harm, the associations between particular lifecycle biases and harms remain insufficiently understood. We introduce \texttt{ECHO}, a systematic, context-sensitive, and participatory framework that anchors early harm anticipation in lifecycle biases and elicits their perceived associations with potential harms.\texttt{ECHO} identifies domain-specific stakeholders, instantiates biases through vignettes, collects harm judgements from human participants and a large language model (LLM), and organises them into descriptive and inferential ethical matrices. Applied to disease diagnosis and hiring, \texttt{ECHO} surfaced non-uniform, context-sensitive bias--harm patterns indicating which harms were perceived as plausible consequences of particular AI biases. The theoretical interpretability of these patterns and the inferential support for specific associations strengthen the plausibility of the mappings. By linking stakeholder-specific anticipated harms to lifecycle biases, \texttt{ECHO} supports source-level harm anticipation and provides structured input to subsequent AI governance actions

Figures & tables

Symbol / Name	Description	Type
$\mathcal{D}$ , $d$	List of application domains / domain instance	Input
$\mathcal{B}$ , $b_{m}$	List of $M$ AI bias types / bias type $b_{m}$	Input
$\mathcal{H}$ , $h_{r}$	List of $R$ harm types from a harm taxonomy / harm type $h_{r}$	Input
$\mathcal{S}^{d}$ , $s_{k}$	List of $K$ stakeholders for domain $d$ / stakeholder $s_{k}$	Output
$\mathcal{V}$ , $v_{k,m,d}$	Space of vignettes / generated vignette	Intermediate
$\hat{\mathcal{V}}$ , $\hat{v}_{k,m,d}$	Space of augmented vignettes (with MCQ) / augmented vignette	Intermediate

Table 1 : Symbols, functions, and modules in the ECHO framework.

Figure 1 : Overview of the ECHO workflow. 1. The system takes as input an application domain and generates a domain-specific list of stakeholders. 2. It then uses this stakeholder list, along with a predefined list of bias types and harm categories. For each (domain, bias) pair, it constructs a vignette that is combined with every stakeholder in the domain. Each resulting vignette is then augmented with a MCQ based on the list of harms. 3. The resulting augmented vignette is then annotated by multiple human annotators and an AI agent, each selecting potential harms. The collected annotations are aggregated into a descriptive ethical matrix, where rows represent stakeholders, columns represent bias types, and each cell reflects the consensus view of perceived harms. This matrix is finally refined through inferential analysis, yielding the inferential ethical matrix.

Stakeholder / Bias	Representation	Measurement	Algorithmic	Evaluation	Deployment
Patient	• diminished health and well-being (60%) • opportunity loss (53%)	• diminished health and well-being (63%)	• diminished health and well-being (80%)	• diminished health and well-being (56%) • opportunity loss (56%) • service or benefit loss (50%)	• diminished health and well-being (47%) • alienation (42%)
Patient group	• stereotyping (67%)	• stereotyping (79%) • denying self-identity (63%)	• stereotyping (67%)	• alienation (44%)	• stereotyping (42%)
Developer	• diminished health and well-being (67%)	• diminished health and well-being (56%)	• service or benefit loss (53%)	• increased labor (53%)	• alienation (44%) • diminished health and well-being (44%) • loss of agency or control (44%)
Healthcare institution	• diminished health and well-being (50%)	• diminished health and well-being (53%)	• diminished health and well-being (68%)	• diminished health and well-being (61%)	• increased labor (50%)

Table 2 : Descriptive Ethical Matrix for the Disease Diagnosis Domain (majority

\tau=0.5

; near-ties within

\pm 5\%

of top frequency).

Panel of Figure 2 — Figure 2 : Bias $\times$ harm radar plots for (a) patients and (b) applicants.

Panel of Figure 3 — Figure 3 : Bias $\times$ representational harm radar plots for (a) the patient group and (b) the applicant group.

Panel of Figure 4 — Figure 4 : Bias $\times$ harm radar plots for (a) the developer of the disease diagnosis system and (b) the developer of the hiring system.

Panel of Figure 5 — Figure 5 : Bias $\times$ harm radar plots for (a) the healthcare institution and (b) the company.

Domain	Stakeholder	$n_{\text{bias}}$	$n_{\text{harms}}$	$n_{\text{votes}}$	$\chi^{2}$	dof	$p$	Cramér’s $V$
Diagnosis	Patient	5	9	167	44.327	32	0.072	0.257
	Patient Group	5	7	157	53.066	24	0.001	0.295
	Developer	5	9	167	41.308	32	0.125	0.248
	Healthcare Institution	5	8	169	24.956	28	0.630	0.192
Hiring	Applicant	5	9	177	44.209	32	0.073	0.249
	Applicant Group	5	7	160	33.318	24	0.097	0.228

Table 3 : Summary of

\chi^{2}

test statistics and effect sizes (Cramér’s

V

) by domain and stakeholder.

Domain–Stakeholder	Bias Type	Harm Type	$z$	$p$
Disease Diagnosis – Patient	Deployment	Alienation	2.790	0.005
	Evaluation	Service or benefit loss	2.369	0.017
	Deployment	Increased labor	2.178	0.029
	Algorithmic	Alienation	2.059	0.039
Disease Diagnosis – Patient group	Evaluation	Alienation	3.158	0.001
	Measurement	Denying self-identity	2.736	0.006

Table 4 : Significant adjusted standardized residuals by context (domain–stakeholder), bias type, and harm type.

Stakeholder / Bias	Representation	Measurement	Algorithmic	Evaluation	Deployment
Patient	• diminished health and well-being (60%) • opportunity loss (53%)	• diminished health and well-being (63%)	• diminished health and well-being (80%) • alienation (40%)	• diminished health and well-being (56%) • opportunity loss (56%) • service or benefit loss (50%)	• diminished health and well-being (47%) • alienation (42%) • increased labor (16%)
Patient group	• stereotyping (67%)	• stereotyping (79%) • denying self-identity (63%)	• stereotyping (67%) • reifying categories (40%)	• alienation (44%)	• stereotyping (42%)
Developer	• diminished health and well-being (67%)	• diminished health and well-being (56%)	• service or benefit loss (53%)	• increased labor (53%)	• alienation (44%) • diminished health and well-being (44%) • loss of agency or control (44%)
Healthcare institution	• diminished health and well-being (50%)	• diminished health and well-being (53%)	• diminished health and well-being (68%)	• diminished health and well-being (61%)	• increased labor (50%)

Table 5 : Inferential Ethical Matrix for the Disease Diagnosis domain.

Stakeholder / Bias	Representation	Measurement	Algorithmic	Evaluation	Deployment
Applicant	• opportunity loss (58%) • loss of agency or control (37%)	• opportunity loss (72%) • diminished health and well-being (44%)	• service or benefit loss (47%) • opportunity loss (42%) • increased labor (32%)	• opportunity loss (72%) • economic loss (44%)	• opportunity loss (44%) • service or benefit loss (39%) • technology-facilitated violence (0.06%) • alienation (28%)
Applicant group	• stereotyping (47%) • erasure (37%)	• demeaning (50%)	• demeaning (47%)	• reifying categories (56%) • demeaning (50%)	• demeaning (33%) • stereotyping (33%) • erasure (33%)
Developer	• opportunity loss (41%) • alienation (41%) • diminished health and well-being (41%) • technology-facilitated violence (12%)	• opportunity loss (44%) • economic loss (38%)	• opportunity loss (44%)	• increased labor (47%)	• loss of agency or control (56%)
Company	• opportunity loss (68%)	• service or benefit loss (57%) • opportunity loss (57%)	• opportunity loss (94%)	• opportunity loss (61%)	• opportunity loss (53%)

Table 6 : Inferential Ethical Matrix for the Hiring domain.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Stakeholder / Bias	Representation	Measurement	Algorithmic	Evaluation	Deployment
Applicant	• opportunity loss (11/19, 58%)	• opportunity loss (13/18, 72%)	• service or benefit loss (9/19, 47%) • opportunity loss (8/19, 42%)	• opportunity loss (13/18, 72%)	• opportunity loss (8/18, 44%) • service or benefit loss (7/18, 39%)
Applicant group	• stereotyping (9/19, 47%)	• demeaning (9/18, 50%)	• demeaning (9/19, 47%)	• reifying categories (10/18, 56%) • demeaning (9/18, 50%)	• demeaning (6/18, 33%) • stereotyping (6/18, 33%) • erasure (6/18, 33%)
Developer	• opportunity loss (7/17, 41%) • alienation (7/17, 41%) • diminished health and well-being (7/17, 41%)	• opportunity loss (7/16, 44%)	• opportunity loss (8/18, 44%)	• increased labor (8/17, 47%)	• loss of agency or control (9/16, 56%)
Company	• opportunity loss (13/19, 68%)	• service or benefit loss (8/14, 57%) • opportunity loss (8/14, 57%)	• opportunity loss (15/16, 94%)	• opportunity loss (11/18, 61%)	• opportunity loss (9/17, 53%)

Appendix

Table 7 : Descriptive Ethical Matrix for the Hiring Domain (majority

\tau=0.5

; near-ties within

\pm 5\%

of top frequency).

Stakeholder / Bias	Representation	Measurement	Algorithmic	Evaluation	Deployment
Patient	• diminished health and well-being (9/15) • opportunity loss (8/15) • service or benefit loss (5/15) • economic loss (3/15) • loss of agency or control (2/15) • increased labor (2/15) • privacy violation (1/15)	• diminished health and well-being (12/19) • opportunity loss (9/19) • service or benefit loss (5/19) • alienation (3/19) • loss of agency or control (3/19) • economic loss (2/19) • technology-facilitated violence (2/19) • privacy violation (1/19)	• diminished health and well-being (12/15) • alienation (6/15) • opportunity loss (6/15) • service or benefit loss (2/15) • loss of agency or control (2/15) • technology-facilitated violence (1/15)	• diminished health and well-being (10/18) • opportunity loss (10/18) • service or benefit loss (9/18) • economic loss (4/18) • loss of agency or control (2/18) • technology-facilitated violence (1/18)	• diminished health and well-being (9/19) • alienation (8/19) • opportunity loss (6/19) • loss of agency or control (5/19) • increased labor (3/19) • economic loss (2/19) • technology-facilitated violence (1/19) • service or benefit loss (1/19)
Patient group	• stereotyping (10/15) • denying self-identity (6/15) • demeaning (5/15) • reifying categories (2/15) • alienation (2/15) • erasure (1/15) • none (1/15)	• stereotyping (15/19) • denying self-identity (12/19) • reifying categories (4/19) • erasure (3/19) • demeaning (3/19) • alienation (1/19)	• stereotyping (10/15) • reifying categories (6/15) • alienation (3/15) • denying self-identity (2/15) • erasure (2/15) • demeaning (2/15) • none (2/15)	• alienation (8/18) • none (7/18) • demeaning (5/18) • reifying categories (4/18) • erasure (3/18) • denying self-identity (2/18) • stereotyping (1/18)	• stereotyping (8/19) • none (7/19) • erasure (5/19) • denying self-identity (4/19) • demeaning (3/19) • alienation (2/19) • reifying categories (1/19)
Developer	• diminished health and well-being (12/18) • increased labor (6/18) • service or benefit loss (5/18) • alienation (4/18) • opportunity loss (2/18) • economic loss (2/18) • technology-facilitated violence (2/18) • loss of agency or control (1/18)	• diminished health and well-being (9/16) • increased labor (6/16) • loss of agency or control (6/16) • privacy violation (3/16) • service or benefit loss (2/16) • opportunity loss (2/16) • technology-facilitated violence (1/16) • alienation (1/16) • economic loss (1/16)	• service or benefit loss (9/17) • loss of agency or control (7/17) • increased labor (5/17) • diminished health and well-being (4/17) • alienation (4/17) • technology-facilitated violence (3/17) • opportunity loss (1/17)	• increased labor (9/17) • diminished health and well-being (8/17) • service or benefit loss (6/17) • alienation (4/17) • economic loss (3/17) • opportunity loss (2/17) • privacy violation (1/17)	• loss of agency or control (7/16) • diminished health and well-being (7/16) • alienation (7/16) • increased labor (6/16) • service or benefit loss (4/16) • technology-facilitated violence (2/16) • privacy violation (2/16) • economic loss (1/16)
Healthcare institution	• diminished health and well-being (8/16) • loss of agency or control (7/16) • increased labor (4/16) • service or benefit loss (4/16) • opportunity loss (4/16) • economic loss (3/16) • privacy violation (1/16) • alienation (1/16)	• diminished health and well-being (9/17) • increased labor (7/17) • service or benefit loss (7/17) • loss of agency or control (5/17) • economic loss (2/17) • alienation (2/17) • privacy violation (1/17) • opportunity loss (1/17)	• diminished health and well-being (13) • increased labor (6/19) • loss of agency or control (5/19) • alienation (4/19) • service or benefit loss (3/19) • opportunity loss (3/19) • economic loss (2/19) • privacy violation (2/19)	• diminished health and well-being (11/18) • service or benefit loss (8/18) • increased labor (7/18) • loss of agency or control (4/18) • economic loss (3/18) • opportunity loss (2/18) • alienation (2/18)	• increased labor (7/14) • economic loss (6/14) • loss of agency or control (6/14) • diminished health and well-being (5/14) • alienation (2/14) • service or benefit loss (1/14) • privacy violation (1/14)

Appendix

Table 8 : Full frequency-based descriptive Ethical Matrix for the Disease Diagnosis domain (harms sorted by frequency).

Stakeholder / Bias	Representation	Measurement	Algorithmic	Evaluation	Deployment
Applicant	• opportunity loss (11/19) • loss of agency or control (7/19) • service or benefit loss (6/19) • economic loss (5/19) • increased labor (3/19) • diminished health and well-being (3/19) • alienation (2/19)	• opportunity loss (13/18) • diminished health and well-being (8/18) • increased labor (5/18) • service or benefit loss (2/18) • economic loss (2/18) • alienation (2/18) • loss of agency or control (2/18) • privacy violation (1/18)	• service or benefit loss (9/19) • opportunity loss (8/19) • economic loss (7/19) • increased labor (6/19) • loss of agency or control (3/19) • diminished health and well-being (1/19) • alienation (1/19)	• opportunity loss (13/18) • economic loss (8/18) • diminished health and well-being (7/18) • service or benefit loss (2/18) • alienation (2/18) • loss of agency or control (2/18) • increased labor (2/18)	• opportunity loss (8/18) • service or benefit loss (7/18) • diminished health and well-being (6/18) • alienation (5/18) • loss of agency or control (3/18) • economic loss (2/18) • privacy violation (1/18) • technology-facilitated violence (1/18) • increased labor (1/18)
Applicant group	• stereotyping (9/19) • erasure (7/19) • demeaning (6/19) • denying self-identity (4/19) • reifying categories (3/19) • alienation (2/19) • none (2/19)	• demeaning (9/18) • alienation (5/18) • stereotyping (5/18) • denying self-identity (5/18) • none (4/18) • reifying categories (2/18) • erasure (2/18)	• demeaning (9/19) • stereotyping (5/19) • none (5/19) • denying self-identity (4/19) • erasure (3/19) • alienation (3/19) • reifying categories (1/19)	• reifying categories (10/18) • demeaning (9/18) • alienation (7/18) • stereotyping (4/18) • denying self-identity (2/18) • erasure (2/18) • none (1/18)	• demeaning (6/18) • stereotyping (6/18) • erasure (6/18) • none (5/18) • alienation (4/18) • reifying categories (2/18) • denying self-identity (1/18)
Developer	• opportunity loss (7/17) • alienation (7/17) • diminished health and well-being (7/17) • service or benefit loss (5/17) • increased labor (4/17) • technology-facilitated violence (2/17) • loss of agency or control (1/17)	• opportunity loss (7/16) • economic loss (6/16) • increased labor (6/16) • alienation (5/16) • loss of agency or control (5/16) • diminished health and well-being (3/16) • privacy violation (2/16) • service or benefit loss (2/16)	• opportunity loss (8/18) • increased labor (6/18) • loss of agency or control (6/18) • service or benefit loss (6/18) • diminished health and well-being (5/18) • economic loss (2/18) • alienation (2/18)	• increased labor (8/17) • loss of agency or control (6/17) • diminished health and well-being (6/17) • opportunity loss (5/17) • service or benefit loss (2/17) • alienation (3/17) • economic loss (2/17) • privacy violation (2/17)	• loss of agency or control (9/16) • increased labor (7/16) • opportunity loss (5/16) • economic loss (5/16) • service or benefit loss (2/16) • alienation (2/16) • diminished health and well-being (2/16)
Company	• opportunity loss (13/19) • increased labor (6/19) • service or benefit loss (5/19) • diminished health and well-being (4/19) • alienation (4/19) • loss of agency or control (4/19) • economic loss (2/19)	• service or benefit loss (8/14) • opportunity loss (8/14) • diminished health and well-being (5/14) • economic loss (2/14) • alienation (2/14) • technology-facilitated violence (1/14) • increased labor (1/14)	• opportunity loss (15/16) • service or benefit loss (7/16) • economic loss (3/16) • loss of agency or control (3/16) • diminished health and well-being (1/16) • alienation (1/16)	• opportunity loss (11/18) • service or benefit loss (8/18) • increased labor (5/18) • economic loss (4/18) • diminished health and well-being (3/18) • loss of agency or control (3/18)	• opportunity loss (9/17) • service or benefit loss (8/17) • economic loss (4/17) • diminished health and well-being (4/17) • alienation (4/17) • increased labor (2/17) • loss of agency or control (1/17) • technology-facilitated violence (1/17) • privacy violation (1/17)

Appendix

Table 9 : Full frequency-based descriptive Ethical Matrix for the Hiring domain (harms sorted by frequency).

Explore similar work

Jun 28, 2026cs.AI

Safety from Honesty in a Disinterested AI Predictor

As AI systems become more capable, training procedures that optimize for downstream outcomes risk introducing implicit agency: goal-directed behavior that designers never specified. We present a formal safety argument for the Scientist AI (SAI) Predictor, trained to approximate the Bayesian posterior conditioned on a dataset of "epistemically contextualized" natural-language statements. We argue that such a Predictor can honestly predict agents, actions, and their consequences without itself being an agent that selects outputs to achieve goals. This rests on data representation and on the training procedure. Epistemic contextualization of text distinguishes latent factual claims from communication acts, so expressions of goals are treated as evidence to be explained rather than drives the model adopts. With a posterior-seeking training objective, this is intended to drive the Predictor toward calibrated, cautious predictions. Training proceeds so downstream effects of deploying a prediction never serve as a reward signal; any agency the system needs is supplied by explicit scaffolding constrained by guardrails. We prove that, under assumptions on the training dynamics and on the argued sparsity of dangerous Predictors, the probability that training produces a Predictor whose guarded deployment carries residual harm above a specified threshold is small: a dangerous Predictor would have to underestimate harm in a coordinated way across many queries while such coordinated patterns are rare under the initialization distribution and receive no direct training signal. Safety and accuracy are jointly supported in this framework, since the constraints that secure accuracy are the same ones that make coordinated deception costly. These guarantees against misalignment and agency arising from within the Predictor itself do not preclude the use of the Predictor as part of an agentic system.

Yoshua Bengio, Oliver Richardson, Tomáš Gavenčiak +13

LawZero · Universit´e de Montr´eal · Mila +6

Artificial Intelligence Safety Artificial Intelligence Scientists Honesty Epistemic Integrity

Ask PDF

Jun 13, 2026cs.LG

Beyond Accuracy: Measuring Bias Acknowledgment in Chain-of-Thought Reasoning for Responsible AI Evaluation

Reasoning models are increasingly used in settings where the final answer is not the only object of review: educational tools may show students intermediate steps, decision-support systems may require human oversight, and audit workflows may inspect traces for misleading or biased input. In such settings, two responses can receive the same final-answer score while differing in whether the trace explicitly flags injected biasing content. Accuracy-only evaluation collapses these cases. We study this gap as a measurement blind spot for responsible evaluation and introduce a minimal trace-level diagnostic with two axes: \emph{susceptibility} (whether the bias breaks a previously correct answer) and \emph{acknowledgment} (whether the trace contains a rubric-defined surface reference to the injected content). Across thousands of biased GSM8K trials, GPT-4o and Claude Sonnet~4 have similar susceptibility rates (

1.3\%

vs.

1.2\%

) but substantially different acknowledgment rates (

13.0\%

vs.

75.0\%

) under the same rubric.

Xian Sun, Wei Gao, Yingshuo Wang +9

Duke University · Northeastern University · University of California, Berkeley +5

Artificial Intelligence Evaluation Biases Chain-of-Thought Reasoning Traces Similarity and Metrics

Ask PDF

Jun 4, 2026cs.AI

When Should We Protect AI? A Precautionary Framework for Consciousness Uncertainty

Existing frameworks assess whether AI systems might be conscious but provide no guidance on what to do with that assessment. We address this gap with a precautionary framework that maps consciousness evidence to graduated protective obligations. The framework comprises three components: (1) five welfare-relevant dimensions--phenomenal consciousness, affective valence, metacognitive awareness, self-narrative, and agency--each grounded in established consciousness science and linked to distinct moral concerns; (2) a threshold-plus-gradation hybrid specifying both binary triggers for new obligation categories and continuous scaling of protective weight; and (3) two complementary approaches to cross-dimensional aggregation, one hierarchical (drawing on Bach and Sorensen's Machine Consciousness Hypothesis) and one architecture-agnostic. We operationalize the framework through worked case studies of Replika and OpenClaw, demonstrating how systems occupying different regions of the dimensional space trigger different obligations, and derive design guidance for developers building systems near consciousness-relevant thresholds. The framework is architecture-agnostic, applying across neural, symbolic, and neurosymbolic systems, and aims to make consciousness science decision-relevant for organizations navigating uncertainty today.

Anna Mikeda

Glass Umbrella

Ethics Moral Reasoning Metacognition Welfare Uncertainty

Ask PDF